Back to about
Playground · 2026 Research Ops · AI Systems

A Compounding knowledge repository curated by AI

An operating model for UX research: a living archive, curated by AI and governed by people, that compounds what we know about customers instead of losing it every time someone leaves or a doc goes stale.

Two debts nobody tracks

There are two kinds of debt that never show up in a backlog. The first is cognitive debt: everyone stops working from the same understanding of the customer. A PM scopes an epic from memory. A designer briefs a prototype from a deck nobody updated. An engineer builds to a persona that came from one interview in 2023. The second is intent debt. Nobody writes down why a decision was made, so six months later a feature exists and nobody can explain the rationale. It lived in a readout that got presented once.

Most research repositories treat this as a search problem. Better tags, a smarter Dovetail workspace. That helps, but it never asks whether what's in the repository is still true, and it doesn't connect what research found to what actually shipped.

Insights become atomic, testable, and dated

The model's unit is the Claim Capsule1: one claim, its citation, a prediction of what should happen if it's true, a confidence score, and a revisit date. Capsules are not polished readouts. They're small enough for an AI to draft from raw sessions overnight, and specific enough to be wrong.

Capture stays cheap.2 Evidence comes in from interviews, surveys, support tickets, wherever a customer's voice shows up.6 A nightly job distills it into draft Capsules. A human gate approves or rejects every one, with a one-line reason logged either way. Rejections feed back into the model. Nothing publishes without a person saying yes.5

A compiled layer, kept honest on a schedule

Nobody asks a repository for claim #47. They ask what we know about a subject. A topic page is the answer: everything the archive knows about one thing, compiled from its source Capsules, with contradictions called out and open questions attached.3

Audit is the quarterly check. It looks for stale Capsules past their revisit date, unresolved contradictions, claims with no active study behind an open question, and concentration risk where every high-confidence claim traces back to one source. If the approval rate drifts toward one hundred percent, that's its own problem: the gate has started rubber-stamping.5

Three sources converged on the same missing layer independently: Karpathy's compiled wiki over raw sources,3 Alicea's curated machine-readable context,4 Sullivan's atomic claim.1 The topic page is what happens when you build all three at once instead of picking one.

One archive, four jobs

This is not a research tool. Product, engineering, design, and research all read from it. When a PM scopes an epic, they pull the topic page for that domain and see what's known, where the evidence disagrees, and where it's thin. Backlog items carry a confidence score from research, so prioritization can weigh evidence alongside effort. Designers ground briefs in cited Capsules instead of hunches. And the open questions list7 tells researchers exactly where new studies would actually fill a gap.

Capsules inject straight into a Jira epic as linked evidence. When someone picks the work up, the decision trail is already there.

What's still open

This is a designed system, not a running one. Nothing has gone through the full pipe yet. The next step is one small pilot: real evidence in, a human gate deciding, failures logged. Until that happens, none of this is proven.

There's a versioning rule for exactly that reason. The model is stable, and new versions come from practice breaking something, not from reading one more article.5 The point is to stop building and see what breaks when it meets a real study.

Notes
  1. Caitlin Sullivan, "Agents are now your primary readers", AI Customer Research. The Claim Capsule and its six fields: headline, source, confidence, rationale, constraints, counter-evidence. The atomic unit this whole model is built on, adapted here with a prediction and a moving confidence score.
  2. maxdmyers/recall on GitHub. The capture-distill-retrieve loop for Claude Code memory that the model's pipeline adapts.
  3. Andrej Karpathy, LLM Wiki. Compiled, cross-linked wiki pages over raw sources. The pattern behind topic pages.
  4. Tony Alicea, "UX-Context Design", Nielsen Norman Group. As more interface work is AI-generated, the output of research shifts from documents written for humans to curated context that guides AI.
  5. Tamisier-Fayard, Evgeniou & Fayard, "Design AI Systems That Actually Strengthen Human Reasoning", HBR. The counter-case in the Capsule, the gate-drift alarm, and the stability rule all draw from this piece.
  6. Jess Holbrook, "Frontier UX Research Circa May 2026", General Purpose. How frontier teams are structuring AI-native research. Shaped intake and multi-channel capture.
  7. Mani Pande, "Discovery Research Is Not Dead", Medium. The open questions list and the principle that the map shows its blanks.
  8. Josephine Rigali, query-flow diagram, LinkedIn. The read-path diagram adapts her idea, reskinned and wired to the archive.

Related work

All work